Rethinking Sparse Autoencoders: Select-and-Project for Fairness and Control from Encoder Features Alone An encoder-centric approach to test-time debiasing with Sparse Autoencoders. Paper